WikiAsp: A Dataset for Multi-domain Aspect-based Summarization

TL;DR

WikiAsp is a multi-domain aspect-based summarization dataset using Wikipedia articles for aspect annotation.

cs.CL 🟡 Intermediate 2020-11-16 10 views
Hiroaki Hayashi Prashant Budania Peng Wang Chris Ackerson Raj Neervannan Graham Neubig
multi-domain aspect-based summarization Wikipedia dataset NLP

Key Findings

Methodology

This study introduces the WikiAsp dataset, utilizing Wikipedia articles' section titles and boundaries as aspect annotations. The research employs automatic extraction, curation, and filtering steps to construct aspect sets and proposes a two-stage baseline model involving aspect identification and summarization.

Key Results

  • The baseline model shows high recall in aspect identification but low precision, indicating potential yet needing improvement.
  • In aspect summarization, the PreSumm model outperforms TextRank in most domains, with an average ROUGE-1 score of 22.94.
  • The extractive Oracle model demonstrates the dataset's potential upper bound, with a ROUGE-1 score of 38.95.

Significance

This research advances open-domain aspect-based summarization by providing a multi-domain dataset. It highlights challenges existing summarization models face in handling multi-domain tasks, such as pronoun handling of quoted sources and consistent explanation of time-sensitive events.

Technical Contribution

Technical contributions include introducing a new multi-domain dataset, WikiAsp, and proposing a two-stage baseline model combining aspect identification and summarization. This dataset provides rich training and testing resources for multi-domain summarization tasks.

Novelty

This is the first to use Wikipedia's section structure for multi-domain aspect annotation, offering an open-domain summarization task framework, contrasting with previous single-domain-focused research.

Limitations

  • The model shows low precision in aspect identification, affecting summary quality.
  • The dataset's construction relies on Wikipedia's section structure, which may not apply to other text types.

Future Work

Future work could explore more refined aspect identification techniques to improve model precision and expand the dataset to cover more domains and text types.

AI Executive Summary

WikiAsp is a dataset designed for multi-domain aspect-based summarization, using Wikipedia articles' section titles and boundaries as aspect annotations. Existing aspect-based summarization models often focus on single domains like product or restaurant reviews, whereas WikiAsp covers 20 different domains, providing an open-domain summarization task framework.

The study proposes a two-stage baseline model, first labeling sentences through aspect identification, then performing aspect summarization. Experimental results indicate challenges existing models face in handling multi-domain tasks, such as pronoun handling of quoted sources and consistent explanation of time-sensitive events.

The contribution of this research lies in advancing open-domain aspect-based summarization and providing rich training and testing resources for multi-domain summarization tasks. Future work could explore more refined aspect identification techniques to improve model precision and expand the dataset to cover more domains and text types.

Deep Analysis

Background

Aspect-based summarization is a task of generating focused summaries based on specific points of interest, widely used for quickly understanding reviews or opinions. However, due to large differences in aspect types across domains, existing models have been domain-specific.

Core Problem

Existing aspect-based summarization models focus on single domains and struggle with multi-domain tasks. Different domains have significantly different aspect types and characteristics, posing challenges for existing models.

Innovation

The study is the first to use Wikipedia's section structure for multi-domain aspect annotation, providing an open-domain summarization task framework. A two-stage baseline model is proposed, combining aspect identification and summarization.

Methodology

  • �� Use Wikipedia articles' section titles and boundaries for aspect annotation
  • �� Construct aspect sets through automatic extraction, curation, and filtering
  • �� Propose a two-stage baseline model: aspect identification and summarization

Experiments

Experiments use the WikiAsp dataset, covering 20 domains. Baseline models include aspect identification and summarization, comparing TextRank and PreSumm.

Results

PreSumm outperforms TextRank in most domains, with an average ROUGE-1 score of 22.94. The extractive Oracle model shows the dataset's potential upper bound, with a ROUGE-1 score of 38.95.

Applications

The WikiAsp dataset can be used for training and testing multi-domain aspect-based summarization models, advancing open-domain summarization research.

Limitations & Outlook

The model shows low precision in aspect identification, affecting summary quality. The dataset's construction relies on Wikipedia's section structure, which may not apply to other text types.

Plain Language Accessible to non-experts

Imagine you're in a library, and each book has different chapters. WikiAsp is like a librarian helping you quickly find the chapters you're interested in. It uses Wikipedia articles' section titles and boundaries to help you find relevant information across different domains.

ELI14 Explained like you're 14

Hey, imagine you're at your school library with a huge encyclopedia. Each topic has many chapters like 'Early Life', 'Career', etc. WikiAsp is like a smart helper that quickly finds the chapters you want to read! It uses Wikipedia's chapters to label different topics, so you can find what you want faster. Isn't that cool?

Glossary

Aspect-based Summarization

A task of generating focused summaries based on specific points of interest, aiding quick understanding of different angles in text.

Used in the paper to describe the goal of the WikiAsp dataset.

WikiAsp

A multi-domain aspect-based summarization dataset using Wikipedia articles' section structure for annotation.

A new dataset proposed in the paper for multi-domain aspect-based summarization research.

PreSumm

A Transformer-based abstractive summarization model using fine-tuned BERT as the source encoder.

Used in experiments to compare with TextRank as a summarization model.

ROUGE Score

A metric for evaluating the quality of text summaries, based on overlap between generated and reference summaries.

Used to evaluate model performance on the WikiAsp dataset.

TextRank

A graph-based ranking model for extracting important sentences.

Used as one of the baseline models in the experiments.

Open Questions Unanswered questions from this research

  • 1 How to improve aspect identification precision to better support multi-domain summarization tasks?
  • 2 How to expand the WikiAsp dataset to cover more domains and text types?

Applications

Immediate Applications

Multi-domain Text Analysis

Researchers can use the WikiAsp dataset for multi-domain text analysis, exploring aspect summarization features across different domains.

Long-term Vision

Open-domain Summarization System

Develop an open-domain summarization system capable of handling multi-domain texts, enhancing information retrieval efficiency.

Abstract

Aspect-based summarization is the task of generating focused summaries based on specific points of interest. Such summaries aid efficient analysis of text, such as quickly understanding reviews or opinions from different angles. However, due to large differences in the type of aspects for different domains (e.g., sentiment, product features), the development of previous models has tended to be domain-specific. In this paper, we propose WikiAsp, a large-scale dataset for multi-domain aspect-based summarization that attempts to spur research in the direction of open-domain aspect-based summarization. Specifically, we build the dataset using Wikipedia articles from 20 different domains, using the section titles and boundaries of each article as a proxy for aspect annotation. We propose several straightforward baseline models for this task and conduct experiments on the dataset. Results highlight key challenges that existing summarization models face in this setting, such as proper pronoun handling of quoted sources and consistent explanation of time-sensitive events.

cs.CL